What Are Connectors and Plugins in ChatGPT and Claude? Bringing Your App Inside the Chat
Connectors and plugins let ChatGPT and Claude use your app directly: read files, run actions and show interactive widgets in the chat. Here is what they are, how they work and why every product should have one.
The way people look for answers has changed
A few years ago, solving a problem online meant opening a search engine, typing some keywords, opening five tabs, comparing results and finally landing on the website or app that could actually do the job.
Today a growing number of people skip most of that. They open ChatGPT or Claude and simply describe what they want:
“Book a table for four at an Italian restaurant near me tomorrow at 8 pm.”
The assistant understands the request, asks a follow-up question if needed and gives an answer. For many users the chat window has become the first place they go for writing, research, planning, coding and everyday tasks.
That raises an important question for every product team:
If your users already live inside a chat window, why make them leave it to use your app?
This is exactly the problem that connectors and plugins solve.
What is a connector?
A connector is a bridge between an AI assistant and an external app or service. It gives the assistant a set of well-described tools it can call on behalf of the user, such as:
- “Search for available restaurants”
- “Check free tables for a date and time”
- “Make a booking”
- “Create a payment link for the deposit”
- “Send the booking confirmation”
Without a connector, an AI assistant can only talk about your product. With a connector, it can use your product.
Most modern connectors are built on the Model Context Protocol (MCP), an open standard introduced by Anthropic and now supported across the industry, including by OpenAI for ChatGPT. MCP defines a common language between an AI client (ChatGPT, Claude, an IDE and so on) and a server you host. You build your MCP server once and it can be used by multiple AI assistants.
A simplified picture looks like this:
User ──► ChatGPT / Claude ──► Your MCP server (connector) ──► Your app, database, APIs
▲ │
└──── results, widgets ◄───┘
The assistant decides when a tool is needed, your connector decides what happens, and your existing backend keeps doing the real work.
What is a plugin?
The words are often used loosely, but a useful way to think about it is:
- A connector is the live link to your service: the tools, the authentication, the data.
- A plugin is a package that bundles that connector with extra knowledge for the assistant: instructions, skills, workflows and the best way to use your tools.
For example, a plugin for a restaurant booking app can include the connector itself plus a “book a table” skill that tells the assistant the right order of steps: confirm the city, date and party size, search for restaurants, check free tables, collect the guest’s details, take the deposit, then share the confirmation. The connector gives the assistant the ability; the plugin gives it the know-how.
Widgets: interactive UI right inside the chat
Text answers are great, but some tasks need a real interface. You would not want to compare restaurant photos, pick a time slot or fill in a form through plain text alone.
This is where widgets come in. Using the Apps SDK in ChatGPT and the MCP Apps standard, a connector can return a small interactive web page that is rendered directly inside the conversation. The widget can:
- Take input: forms, buttons, dropdowns and editable lists.
- Accept files: the user drops a document or image straight into the widget, or the assistant passes a file the user already attached in the chat.
- Show live progress: status, step indicators and countdowns that update by themselves.
- Call your tools: when the user clicks a button, the widget calls the same connector tools the assistant uses.
- Keep the assistant informed: the widget can tell the model what the user did, so the conversation stays in sync.
Take the restaurant booking example. A widget can guide the user through the whole flow without leaving the chat:
- Choose: browse a few matching restaurants as cards with photos, ratings and prices.
- Pick a slot: tap one of the free time slots for the chosen date.
- Details: fill in the name, phone number and any special requests in a short form.
- Pay: open a secure payment page with one click for the deposit, and use a “check payment status” button to continue.
- Confirm: see the booking confirmation, with buttons to add it to a calendar or get directions.
The user types one sentence and the rest happens in a friendly panel right under the message.
Why connectors and plugins matter
1. The user just needs to type
No sign-up page to find, no menu to learn, no tutorial to watch. The user describes what they want in their own words and in their own language, and the assistant does the navigating.
2. Meet users where they already are
Your product shows up inside the tools people already open dozens of times a day. Instead of fighting for a click on a search results page, you become part of the answer.
3. Fewer steps, less drop-off
Every redirect, new tab and login screen loses users. Keeping the flow in one place, chat plus widget, removes most of those breaking points.
4. Files and context come for free
The assistant already knows what the user has said and which files they attached. A connector can receive that file directly, so the user does not have to upload it a second time.
5. Build once, appear everywhere
Because MCP is an open standard, the same server can power a ChatGPT app, a Claude connector and other MCP-compatible clients. Your existing API and business logic stay the same; the connector is a thin, focused layer on top.
6. Smarter help, out of the box
The assistant can explain errors in plain language, suggest the next step, answer “what does this column mean?” and translate on the fly, all without you building a chatbot.
7. A new discovery and growth channel
App directories inside ChatGPT and Claude are a fresh way for people to find your product, often at the exact moment they need it.
What to keep in mind
Bringing your app into a chat window is powerful, but it also means your service is reachable from a new place. A production-ready connector should take care of:
- Security: strict validation of every input, safe file downloads, careful handling of tokens and links, and a locked-down widget.
- Rate limiting and abuse protection: per user and per IP limits, so automated traffic cannot overload your system.
- Clear tool descriptions: the assistant chooses tools based on their descriptions, so they must be precise and honest.
- Graceful fallbacks: not every client supports widgets or file uploads in the same way, so plan a text-only path too.
- Payments outside the sandbox: chat widgets run in a restricted frame, so sensitive steps like payment usually open on your own secure page.
We go deeper into each of these in Part 1, Part 2 and Part 3 of our build series.
Conclusion
The shift from searching to chatting is already here. Connectors give AI assistants the ability to use your product, plugins teach them how to use it well, and widgets let users interact with it visually, all inside a single conversation.
For users it means doing more by just typing. For businesses it means being present exactly where attention is moving.
Stay connected
We have built exactly this for our own product, Bank Statements CSV / Card Statement CSV: a connector and plugin that lets users convert their PDF bank and credit card statements to CSV right inside ChatGPT and Claude. In the upcoming posts, we will share how we planned and built it.
This is the first post in our series on AI connectors and plugins. Next, we show how you can build one for your own use case:
- Part 1: Plan before you code
- Part 2: Build the MCP server, tools and widget
- Part 3: Secure, test and publish on ChatGPT and Claude
New to MCP itself? Read Model Context Protocol (MCP): The New Standard for AI Integration first.